EP3571664B1

EP3571664B1 - Determining the location of a mobile device

Info

Publication number: EP3571664B1
Application number: EP18701218.2A
Authority: EP
Inventors: Ronald Clark; Sen WANG; Niki TRIGONI
Original assignee: Oxford University Innovation Ltd
Current assignee: Oxford University Innovation Ltd
Priority date: 2017-01-23
Filing date: 2018-01-17
Publication date: 2021-05-12
Anticipated expiration: 2038-01-17
Also published as: AU2018209336A1; AU2018209336B2; CN110770758B; US20210407122A1; WO2018134587A1; JP7121017B2; CN110770758A; EP3571664A1; JP2020505695A; US11348274B2

Description

Field of the Invention

The present invention concerns determining the location of a mobile device. More particularly, but not exclusively, the invention concerns using neural networks to determine the location of a mobile device from a combination of images captured by a camera of the mobile device and motion information captured by inertial sensors of the mobile device.
The invention is particularly, but not exclusively, applicable where the mobile device is an autonomous robot. However, the invention is also applicable to other types of mobile and wearable devices, such as mobile phones, smart watches and the like.
"Location" as discussed herein may refer to absolute location, such as that a mobile device is at a position on the Earth defined by latitude and longitude, and may also refer to relative location with respect to another position (e.g. a mobile device is a distance and direction from an initial starting position). Determination of location also frequently includes determination of orientation, in absolute terms with respect to the Earth's magnetic field for example, and in relative terms as rotation of a certain amount with respect to an initial orientation.

Background of the Invention

It is desirable to be able to determine the location of mobile devices where no GPS signal is available. This is particularly the case for autonomous robots, to allow accurate navigation. A known approach is to use images from a camera in order to determine location. However, such systems tend to require very precise calibration of the camera if location is to be reliably determined. Traditional visual odometry techniques include sparse methods (comprising several steps, including detection and matching of features, motion estimation and optimisation), and direct methods (comprising steps of motion estimation and optimisation). Such techniques tend to require precise camera calibration, and often fail in texture-poor environments (i.e. where there are with few features) or when a camera capturing images is rotated rapidly. In addition, while such systems are usually able to determine the shape of a path of travel, they are usually unable to estimate scale, i.e. the actual distance travelled.
Alternatively, it is known to use a neural network to process the images from a camera in order to determine location. Mohanty et al, DeepVO: A Deep Learning approach for Monocular Visual Odometry, arXiv:1611.06069, published 18 November 2016, discloses such a system. However, there are various problems with known systems that use neural networks. They tend to require training for the specific environments in which they will be used, so cannot be used in new environments without suitable training occurring first.
It is known to use neural networks to learn the relation between camera poses and inertial data. Learning to Fuse: A Deep Learning Approach to Visual-Inertial Camera Pose Estimation, IEEE ISMAR, published 15 December 2016, discloses such a system. However, there are various problems with known approaches using neural networks. Disadvantageously, they require the camera data to be processed using traditional methods to obtain accurate pose estimates, and need markers to operate. Furthermore, they cannot be trained end-to-end using raw data as input.
Systems which combine image data and motion data to determine location are known. However, a problem with such systems is the image data and motion data needs to be time-synchronised, making such systems difficult to set up, as inertial sensors are not designed to provide motion data at the same intervals as cameras capture images. Further, any miscalibration of the camera or inertial sensors can lead to significant errors.
The present invention seeks to mitigate the above-mentioned problems. Alternatively and/or additionally, the present invention seeks to provide improved methods of determining the location of a mobile device.

Summary of the Invention

In accordance with a first aspect of the invention there is provided a computer-implemented method of determining the location of a mobile device comprising a camera and at least one inertial sensor, comprising the steps of:

capturing, using the camera, a sequence of images over a period of time;
for pairs of consecutive images from the sequence of images, determining, using a convolutional neural network, features indicative of the motion of the device between the time the first image of the pair of images was captured and the time the second image of the pair of images was captured;
capturing, using the at least one inertial sensor, data indicative of the motion of the device;
determining, using a first recurrent neural network, features indicative of the motion of the device from the data indicative of the motion of the device from the at least one inertial sensor;
for the sequence of consecutive images, determining, using a second recurrent neural network, features indicative of the location of the device from the features determined by the convolutional neural network and the first recurrent neural network; and
for the sequence of consecutive images, determining the location of the device from the features determined by the second recurrent neural network.

By using first and second neural networks to determine features from both the images and the motion data, and then using a third neural network to determine the location from those features, it has been found that much more robust and reliable location determination is possible. In particular, careful calibration of camera and inertial sensor separately, and particularly with respect to each other, as well as time synchronisation between the two, is not required. Further, all the neural networks can be trained simultaneously by training the system as a whole. This means in particular that the first neural network and second neural network can be trained to determine whatever motion features are best for the operation of the system as a whole, rather than being trained to determine motion features with preselected attributes which may not in fact be the optimal type of features to use. Further, the first neural network and second neural network will be trained to determine features that are best in combination with each other, i.e. any synchronisation of the features from the first neural network and second neural network that is desirable will be handled automatically as part of the training. In addition, as the third neural network takes the features from the first neural network and second neural network, this will at the same time be trained to optimally handle any lack of calibration or synchronisation that there may be. (The extent to which the third neural network will be trained to handle any lack of synchronisation will result automatically from the training; in general, if synchronisation leads to better location determination then the first neural network and second neural network will be trained to be synchronised, while if a lack of synchronisation can be effectively handled by the third neural network then it will be trained to do so.)
A convolutional neural network is particularly suited for operating upon data with a large number of parameters such as image data.
Preferably, the orientation of the device is determined in addition to the location of the device. Thus, the "pose" of the device is determined.
Preferably, the images of the sequence of images are monocular images.
The at least one inertial sensor includes an accelerometer and/or a gyroscope. It may additionally or alternatively include a magnetometer, or any other appropriate inertial sensor.
Preferably, the method further comprises the step of determining, from the features determined by the first recurrent neural network, relative location and orientation information for the device for each pair of consecutive captured images.
Advantageously, the method further comprising the step of determining corresponding uncertainty information for the relative location and orientation information for each pair of consecutive captured images.
Advantageously, the method further comprises the step of training the convolutional, first recurrent and second recurrent neural networks to maximise the correlation between the features determined by the convolutional neural network and the features determined by the first recurrent neural network. As the features of the convolutional and the first recurrent neural network are both indicative of the motion of the mobile device, in general correlation between the features should be expected, and so training to maximise such correlation is advantageous.
The device may be an autonomous robot.
In accordance with a second aspect of the invention there is provided a mobile device comprising:

memory;
a processor;
a camera;
at least one inertial sensor;
wherein the device is arranged to:
- capture, using the camera, a sequence of images over a period of time;
- from pairs of consecutive images from the sequence of images, determine, using a convolutional neural network provided by the processor, features indicative of the motion of the device between the time the first image of the pair of images was captured and the time the second image of the pair of images was captured;
- capture, using the at least one inertial sensor, data indicative of the motion of the device;
- determine, using a first recurrent neural network, features indicative of the motion of the device from the data indicative of the motion of the device from the at least one inertial sensor;
- for the sequence of consecutive images, determine, using a second recurrent neural network provided by the processor, features indicative of the location of the device from the features determined by the convolutional neural network and the first recurrent neural network; and
- for the sequence of consecutive images, determine the location of the device from the features determined by the first recurrent neural network.

Preferably, the device is arranged to determine the orientation of the device in addition to the location of the device.
Preferably, the images of the sequence of images are monocular images.
The at least one inertial sensor may include an accelerometer and/or a gyroscope.
Preferably the second neural network is a Long Short-Term Memory neural network. Preferably the third neural network is a Long Short-Term Memory neural network.
Advantageously, when both location and orientation information for the device are determined the device is further arranged to determine corresponding uncertainty information for the relative location and orientation information for each pair of consecutive captured images.
In accordance with a third aspect of the invention there is provided a computer program product arranged, when executed on a mobile device, to perform any of the methods described above.
In accordance with a fourth aspect of the invention there is provided a computer program product arranged, when executed on a mobile device, to provide any of the mobile devices described above.

Description of the Drawings

Embodiments of the present invention will now be described by way of example only with reference to the accompanying schematic drawings of which:

Figure 1 is a schematic diagram of an autonomous robot in accordance with an embodiment of the invention;
Figure 2 is a flowchart showing operation of the autonomous robot of Figure 1 to estimate its location;
Figure 3 is a schematic diagram representing the process of Figure 2; and
Figure 4 is a schematic diagram of an LSTM used in the process of Figures 2 and 3.

Detailed Description

A schematic diagram of an autonomous robot in accordance with an embodiment of the invention is shown in Figure 1. The autonomous robot 1 comprises a processor 2. It will be appreciated that in different embodiments the processor 2 may be a single processor system, a dual processor system, or any other suitable processor system. The processor 2 is in communication with a camera 3, a inertial sensor 5, and a memory 4 which stores (amongst other things) images captured by the camera 3 and motion data captured by the inertial sensor 5.
The inertial sensor 5 may be a gyroscope, accelerometer, magnetometer or any other suitable inertial sensor, or indeed a combination of such devices.
The operation of the autonomous robot 1 to determine its location is now described with reference to the flowchart of Figure 2. The process is also shown schematically in Figure 3. At each time step the autonomous robot 1 determines its location based on the information currently available to it. Figure 3 shows three successive time steps t, t+1 and t+2, where for each time step a "pose" for the autonomous robot 1 is determined, Pose_t, Pose_t+1 and Pose_t+2 respectively, where a pose is a combination of location and heading information (i.e. the direction in which the autonomous robot is facing).
First, a pair of images consecutive captured by the camera 3 are obtained (step 21, portion 31 of Figure 3). Each image is then pre-processed (step 22, portion 32 of Figure 3), by subtracting the mean RGB channel values from an example set of images. The set of images may be those with which the autonomous robot 1 is trained, for example, as described in detail below. In addition, the images are resized to be a multiple of 64. However, such pre-processing is optional, and in other embodiments does not occur. As can be seen in Figure 3, for time step t the initial pair of images results in pre-processed images RGBt and RGB_t+1, while for time step t+1 the pre-processed images are RGB_t+1 and RGB_t+2, and so on for the other time steps.
A convolutional neural network (CNN), implemented by the processor 2 and memory 4, takes the pair of pre-processed images and uses it to determine features (step 23, portion 33 of Figure 3). The CNN determines the features in accordance with its training, which has been performed as described below.
CNNs are a type of neural network that incorporates convolutional layers in the network structure and so, in contrast to fully-connected layers used in other types of neural network, are able to take advantage of the spatial regularity of data. This means that the number of parameters required for the CNN is significantly reduced, allowing them to operate on high-dimensional input (such as raw image data). In CNNs, multiple convolutional operations are applied at each convolutional layer, to determine a number of features from the output map of the previous layer. The filter kernels with which the maps are convolved are learned during training.

The CNN takes as input a tensor generated by stacking the pair of pre-processed consecutive images. The CNN is composed of nine convolutional layers, each except the last followed by a Rectified Linear Unit (ReLU) non-linearity activation, giving 17 layers in total. The layers are configured as follows:

Layer	Receptive field size	Padding	Stride	Number of channels
Conv1	7x7
	3	2	64
Conv2	5x5		2	2	128
Conv3	5x5		2	2	256
Conv3_1	3x3		1	1	256
Conv4	3x3		1	2	512
Conv4_1	3x3		1	1	512
Conv5	3x3		1	2	512
Conv5_1	3x3		1	1	512
Conv6	3x3		1	2	1024

The sizes of the receptive fields in the network gradually reduce from 7 x 7 to 5 x 5 and then 3 x 3, to capture small interesting features. Zero-paddings are introduced either to adapt to the configurations of the receptive fields, or to preserve the spatial dimension of the tensor after convolution. The number of the channels, i.e. the number of filters for feature detection, increases to learn various features.
In the present embodiment the CNN has 55 million trainable weighs, but it will be appreciated that in other embodiments different numbers of weights could be used.
The features from the final layer (i.e. Conv6) are then the output of the CNN.
Simultaneously, a first recurrent neural network (RNN) takes the motion data from the inertial sensor 5, and uses it to determine features (step 24, portion 34 of Figure 3). Similarly to the CNN, the first RNN determines the features in accordance with its training, which has been performed as described below.
RNNs are a type of neural network in which the layers operate on the input, but also on delayed versions of the hidden layers and/or output. In this way, RNNs have an internal state that they can use as "memory" to keep track of past inputs and corresponding decisions.
In the present embodiment, RNNs with a Long Short-Term Memory (LTSM) architecture is used (of which various variations exist), as shown in Figure 4, in which
denotes element-wise product, and ⊕ denotes addition of two vectors. The contents of the memory cell is stored in c_t. The input gate it controls how the input enters into the contents of the memory for the current time-step. The forget gate ft determines when the memory cell should be emptied by producing a control signal 0 to 1 which clears the memory cell as needed. Finally, the output gate o_t determines whether the contents of the memory cell should be used at the current time step. The operation of the RNN is described by the following equations: $i_{t} = σ (W_{xi} x_{t} + W_{hi} h_{t - 1} + W_{ci} c_{t - 1} + b_{i})$
$f_{t} = σ (W_{xf} x_{t} + W_{hf} h_{t - 1} + W_{cf} c_{t - 1} + b_{f})$
$z_{t} = \tanh (W_{xc} x_{t} + W_{hc} h_{t - 1} + b_{c})$
$c_{t} = f_{t} ⊙ c_{t - 1} + i_{t} ⊙ z_{t}$
$o_{t} = act (W_{xo} x_{t} + W_{ho} h_{t - 1} + W_{co} c_{t} + b_{o})$
$h_{t} = o_{t} ⊙ \tanh (c_{t})$
The parameters W_¡,j and b_i fully parameterise the operation of the RNN, and are learned during training. The recurrent hidden layer allows the network to make use of the temporal regularity of the input data to improve its performance.
In the present embodiment the LSTM has two layers with cells of 2000 units, but it will be appreciated that in other embodiments different numbers of layers and units could be used.
The features determined by the CNN and by the first RNN are then passed to a second RNN, which uses them to determine motion features (step 25, portion 35 of Figure 3). In the present embodiment the second RNN takes as input, as well as the hidden state carried over from the previous time step, the pose determined for the previous time step is directly fed to the RNN. This can be seen in Figure 3, in which the pose for a time step is fed to the LSTM box for the next time step. The reason for doing this is because, for location estimation, the output is essentially the integration of successive displacements at each time step. Consequently, the determined pose for the previous time step is of particular significance.
Next, the motion features determined by the RNN (of high dimensionality) are passed to a fully connected layer (step 26), which outputs features of low dimensionality (at least 6 for pose, 6 for uncertainty, and potentially more for each if a mixture of Gaussians is used to estimate pose and uncertainty).
Next, the low dimensional features from the fully connected layer are passed to an SE(3) layer (step 27, SE3 box of portion 34 of Figure 3). The SE(3) composes successive motion features for each time step, in order to at each time step determine a location (in fact a pose, such as Pose_t for time step t) for the autonomous robot 1.
SE3 is the special Euclidian group whose elements are transformational matrices, which consist of a rotation from the special orthogonal group SO3 and a translation vector: $T = \{(\begin{matrix} R & T \\ 0 & 1 \end{matrix}) ∣ R \in SO (3), T \in R^{3}\} .$
Producing transformation estimates belonging to SE3 is not straightforward, as the SO3 component needs to be an orthogonal matrix. However, the Lie Algebra se3 of SE3 can be described by components which are not subject to orthogonality constraints: $\frac{ξ}{dt} = \{(\begin{matrix} {[ω]}_{x} & v \\ 0 & 0 \end{matrix}) ∣ v \in so (3), v \in R^{3}\},$
Conversion between se3 and SE3 can then be accomplished using the exponential map: $\begin{array}{l} \exp : se 3 \to SE 3 \end{array}$
In an alternative embodiment, quaternionic representations of rotations are used instead of the matrix representations. In particular, the ω components are converted to vector: $w_{t} = [0, ω_{r}, ω_{y}, ω_{z}]$
${\dot{q}}_{t} = \frac{1}{2} w_{t} \times q_{t}$
$q_{t} = q_{t} + dt {\dot{q}}_{t}$
Calculating the gradients of these quantities can then be done using only simple linear algebra operations. In addition, the expensive eigenvalue decompensation required to calculate the exponential map is avoided.
Thus, in this way the autonomous robot 1 uses the images from the camera 3 to estimate its location, in particular its pose.
Estimating location from successive sensor measurements (i.e. odometry) is inevitably subject to drift. As a result, it is common to use it in tandem with loop-closure, map-matching or pose-graph optimisation methods to create Simultaneous Localisation and Mapping (SLAM) systems. A key aspect of integrating odometry measurements into such systems is the availability of uncertainty estimates.
In order to provide such estimates, the output of the fully connected layer is used (before the SE(3) layer). Estimates produced by the fully connected layer are compared with ground truth pose information from training data, resulting in a distribution of errors in pose (location and orientation). A max likelihood approach is then used to train the prediction of the mixture of Gaussian distributions that represents the uncertainty.
In order to operate, neural networks of course need to be trained, which is done by providing test data and a cost function to be minimised. The training of the CNN and the first and second RNNs of the autonomous robot 1 as now described; in fact, all are trained simultaneously.
As discussed above, the system of the present embodiment estimates both pose and uncertainty. The test data will be a sequence of images, with "ground truth" poses (i.e. correct poses) for those images. The cost function for training is composed of two parts, the first part relating to the pose estimate and the second part relating to the uncertainty estimate. For pose estimation, the first part of the cost function trains the system to minimise the difference between the estimated poses and the ground truth poses. For uncertainty estimation, the second part of the cost function trains the system by comparing the output of the neural network with pose labels. The training is then done by back-propagation through time, to adjust the weights of the CNN and the first and second RNNs to best minimise the result of the cost function.
In this way, it can be seen that the CNN and the first RNN are both trained to provide the most appropriate features for input to the second RNN, and the second RNN is simultaneously trained to most accurately determine the pose (and uncertainty thereof) of the autonomous robot 1 from those features (and previous determinations). The CNN and first RNN in particular are not trained to best provide features of any particular type or with any specific properties; rather they are simply trained to provide features which are best for the operation of the system as a whole. However, in certain embodiments, to speed up the initial training process, the CNN is initially trained in isolation (or otherwise provided with weights that have the effect of such training) to provide features indicative of movement between successive images. This provides an initial state for the CNN, which is then further optimally trained as the system is trained as a whole.
In a particularly advantageous embodiment, unsupervised training of the neural networks is performed using correlation. In particular, as both the images from the camera 3 and the motion data from the inertial sensor 5 should be indicative of the same movement of the autonomous robot 1, training can be done to maximise correlation between the features determined by the CNN from the images and the features determined by the first RNN from the motion data. In other words, where the images suggest a large movement so should the motion data (and vice versa), and conversely where the images suggest a small movement so should the motion data (and vice versa), so in general there should be correlation between the features from the CNN and the first RNN, and so training to maximise correlation is advantageous.
While the present invention has been described and illustrated with reference to particular embodiments, it will be appreciated by those of ordinary skill in the art that the invention lends itself to many different variations not specifically illustrated herein.
Where in the foregoing description, integers or elements are mentioned which have known, obvious or foreseeable equivalents, then such equivalents are herein incorporated as if individually set forth. Reference should be made to the claims for determining the true scope of the present invention, which should be construed so as to encompass any such equivalents. It will also be appreciated by the reader that integers or features of the invention that are described as preferable, advantageous, convenient or the like are optional and do not limit the scope of the independent claims. Moreover, it is to be understood that such optional integers or features, whilst of possible benefit in some embodiments of the invention, may not be desirable, and may therefore be absent, in other embodiments.

Claims

A computer-implemented method of determining the location of a mobile device comprising a camera and at least one inertial sensor, comprising the steps of:
capturing, using the camera, a sequence of images over a period of time;

for pairs of consecutive images from the sequence of images, determining, using a convolutional neural network, features indicative of the motion of the device between the time the first image of the pair of images was captured and the time the second image of the pair of images was captured;

capturing, using the at least one inertial sensor, data indicative of the motion of the device;

determining, using a first recurrent neural network, features indicative of the motion of the device from the data indicative of the motion of the device from the at least one inertial sensor;

for the sequence of images, determining, using a second recurrent neural network, features indicative of the location of the device from the features determined by the convolutional neural network and the first recurrent neural network; and

for the sequence of images, determining the location of the device from the features determined by the second recurrent neural network.
A method as claimed in claim 1, wherein the orientation of the device is determined in addition to the location of the device.
A method as claimed in claim 1 or 2, wherein the images of the sequence of images are monocular images.
A method as claimed in any preceding claim, wherein the at least one inertial sensor includes an accelerometer and/or a gyroscope.
A method as claimed in any preceding claim, further comprising the step of determining, from the features determined by the first recurrent neural network, relative location and orientation information for the device for each pair of consecutive captured images.
A method as claimed in claim 5, further comprising the step of determining corresponding uncertainty information for the relative location and orientation information for each pair of consecutive captured images.
A method as claimed in any preceding claim, further comprises the step of training the convolutional, first recurrent and second recurrent neural networks to maximise the correlation between the features determined by the convolutional neural network and the features determined by the first recurrent neural network.
A method as claimed in any preceding claim, wherein the device is an autonomous robot.
A mobile device comprising:
memory;

a processor;

a camera;

at least one inertial sensor;
wherein the device is arranged to:
capture, using the camera, a sequence of images over a period of time;

from pairs of consecutive images from the sequence of images, determine, using a convolutional neural network provided by the processor, features indicative of the motion of the device between the time the first image of the pair of images was captured and the time the second image of the pair of images was captured;

capture, using the at least one inertial sensor, data indicative of the motion of the device;

determine, using a first recurrent neural network, features indicative of the motion of the device from the data indicative of the motion of the device from the at least one inertial sensor;

for the sequence of images, determine, using a second recurrent neural network provided by the processor, features indicative of the location of the device from the features determined by the convolutional neural network and the first recurrent neural network; and

for the sequence of images, determine the location of the device from the features determined by the first recurrent neural network.
A device as claimed in claim 9, wherein the device is arranged to determine the orientation of the device in addition to the location of the device.
A device as claimed in claim 9 or 10, wherein the images of the sequence of images are monocular images.
A device as claimed in any of claims 9 to 11, wherein the at least one inertial sensor includes an accelerometer and/or a gyroscope.
A device as claimed in claim 10, further arranged to determine corresponding uncertainty information for the relative location and orientation information for each pair of consecutive captured images.
A computer program product arranged, when executed on a mobile device, to perform the method of any of claims 1 to 8.
A computer program product arranged, when executed on a mobile device, to provide a mobile device as claimed in any of claims 9 to 13.